News Stance Discrimination Based on a Heterogeneous Network of Social Background Information Fusion

Abstract

Media with partisan tendencies publish news articles to support their preferred political parties to guide the direction of public opinion. Therefore, discovering political bias in news texts has important practical significance for national election prediction and public opinion management. Some biased news often has obscure expressions and ambiguous writing styles. By bypassing the language model, the accuracy of methods that rely on news semantic information for position discrimination is low. This manuscript proposes a news standpoint discrimination method based on social background information fusion heterogeneous network. This method expands the judgment ability of creators and topics on news standpoints from external information and fine-grained topics based on news semantics. Multi-attribute features of nodes enrich the feature representation of nodes, and joint representation of heterogeneous networks can reduce the dependence of position discrimination on the news semantic information. To effectively deal with the position discrimination of new news, the design of a multi-attribute fusion heterogeneous network is extended to inductive learning, avoiding the cost of model training caused by recomposition. Based on the Allsides dataset, this manuscript expands the information of its creator's social background and compares the model for discriminating political positions based on news content. In the experiment, the best transductive attribute fusion heterogeneous network achieved an accuracy of 92.24% and a macro F1 value of 92.05%. The effect is improved based purely on semantic information for position discrimination, which proves the effectiveness of the model design.